An underground tunnel intrusion event identification method and system

Abstract

This invention discloses a method and system for identifying intrusion events in underground tunnels. The method includes: acquiring raw acoustic signals from underground tunnels using a distributed fiber optic acoustic sensing system; extracting high-dimensional initial features covering both the time domain and the wavelet packet time-frequency domain; and using a feature selection framework that integrates multiple evaluation criteria to select the most discriminative explicit waveform feature subset. Based on this, a dual-branch classification network is constructed, utilizing multi-scale convolution and channel attention modules to extract implicit temporal features, and aligning the explicit waveform features through feature mapping. The two feature paths are concatenated and fused, and an improved adaptive focus loss function is used to train the network for classification and early warning of underground tunnels. This invention's technical solution effectively constrains the feature space of conventional deep learning models through explicit waveform features, and the implicit feature recognition is more reliable and accurate.

Description

Technical Field

[0001] This invention belongs to the field of mine safety monitoring and dynamic disaster early warning technology, and specifically relates to a method and system for identifying underground roadway intrusion events. Background Technology

[0002] As mining operations extend deeper, shallow roadways are susceptible to geological disasters such as localized rockfalls and abnormal water outbursts due to complex geostress disturbances. Introducing Distributed Acoustic Sensing (DAS) technology can effectively address the limited coverage of traditional point-based monitoring, enabling continuous monitoring of long-distance underground spaces.

[0003] However, the practical application of DAS technology in mine roadways faces multiple technical bottlenecks. On the one hand, the high-dimensional waveform feature space of DAS signals inevitably contains a large amount of redundant interference, and there is still a lack of robust feature selection mechanisms adapted to the complex acoustic environment of mines. On the other hand, traditional machine learning methods have poor generalization ability under complex nonlinear signals, while conventional deep learning models have "black box" characteristics and lack clear physical waveform feature guidance. When faced with transient events with highly similar waveforms, they are prone to cross-class confusion and misjudgment, resulting in a high false alarm rate in the early warning system and seriously affecting the reliability of mine safety early warning. Summary of the Invention

[0004] This invention discloses a method and system for identifying underground tunnel intrusion events. It addresses the problem of high confusion rates in classifying highly similar transient impact events due to severe redundancy in the high-dimensional features of DAS signals and the lack of explicit physical feature guidance in conventional deep learning models. This invention extracts a subset of low-redundancy explicit waveform features using a multi-criteria feature selection framework, constructs a dual-branch classification network integrating implicit deep features and explicit waveform features, and combines an adaptive focus loss function for early warning classification of mine tunnel intrusion events. The technical solution of this invention effectively constrains the feature space of conventional deep learning models through explicit waveform features, solving the technical challenge of easily confusing highly similar transient impact signals under strong background noise. The early warning results are more reliable and accurate than those based on a single implicit feature.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] A method for identifying intrusion events in underground tunnels includes the following steps:

[0007] Step S1: Deploy a distributed fiber optic acoustic wave sensing system in the underground tunnel, including a sensing optical cable laid along the tunnel sidewall and a distributed fiber optic acoustic wave sensing host connected to the sensing optical cable at the tunnel exit. Acoustic signals are collected and time-series signals of preset events are captured through the distributed fiber optic acoustic wave sensing system.

[0008] Step S2: Extract a multi-dimensional waveform feature pool covering the time domain and wavelet packet time-frequency domain features from the time series signal, and use a multi-criteria feature filtering method to filter and reduce the dimensionality of each feature in the multi-dimensional waveform feature pool to obtain a target explicit waveform feature subset;

[0009] Step S3: Construct a dual-branch feature extraction network comprising an implicit deep feature extraction branch and an explicit waveform feature transformation branch; wherein, the implicit deep feature extraction branch receives the time series signal, extracts temporal features at different time-frequency scales in parallel through a multi-scale convolution module, and adaptively weights the feature channels through a channel attention module to suppress environmental noise, outputting implicit deep features; the explicit waveform feature transformation branch receives the target explicit waveform feature subset obtained in step S2, nonlinearly maps the target explicit waveform feature subset to a high-dimensional semantic space through a feature mapping layer, and outputs the mapped explicit waveform features;

[0010] Step S4: Concatenate and fuse the implicit deep features and the mapped explicit waveform features in the channel dimension to form a joint feature vector, and combine the adaptive focus loss function and learning rate scheduling strategy to complete the network model training;

[0011] Step S5: Input the real-time acoustic signal to be tested into the trained network model, and output the category probability of each preset event to realize the classification and early warning of underground tunnel intrusion events.

[0012] Optionally, in the step of using a multi-criteria feature selection method to select and reduce the dimensionality of each feature in the multi-dimensional waveform feature pool, the multi-criteria feature selection method includes evaluation from the following three complementary perspectives:

[0013] (1) Linear discriminant evaluation: The F-value is calculated using analysis of variance as a criterion for original importance score, quantifying the linear discriminant ability of each feature for the classification label of the preset event. The calculation formula is as follows:

[0014] ;

[0015] in, For the number of categories, The total number of samples, For the first Number of samples per class For the first Class feature mean This is the population mean;

[0016] (2) Nonlinear correlation evaluation: Mutual information (MI) is used to quantify the characteristics. To determine the classification label The information gain provided, as a criterion for the original importance score, is calculated using the following formula:

[0017] ;

[0018] in, It is the marginal entropy of the category. It is conditional entropy. It is a joint probability distribution. and It is a marginal probability distribution;

[0019] (3) Global Contribution Evaluation of Tree Model: The contribution of each feature in the ensemble decision is calculated using a random forest and an extreme gradient boosting model. The random forest quantifies the importance of a feature by calculating the decrease in Gini impurity across all decision trees, serving as a criterion for the initial importance score. The calculation formula is as follows:

[0020] ;

[0021] in, is the number of trees, b is the index of a single decision tree, and T b Let b be the set of all nodes of the b-th tree. It is a node The splitting characteristics, Is the node reached The sample proportion This is the decrease in the Gini index caused by the split at this node;

[0022] The extreme gradient boosting model uses the gain from objective function optimization to calculate importance, which serves as the original importance score as a criterion. The calculation formula is as follows:

[0023] ;

[0024] in, and It is the sum of the first derivatives of the left and right child nodes. and It is the sum of the second derivatives. and This is the regularization parameter.

[0025] Optionally, after obtaining the original importance scores of the four criteria, min-max normalization is used to map all scores to the [0,1] interval, as shown in the following formula:

[0026] ;

[0027] in, For the first One feature in The normalized scores under the criteria correspond to the following: , , and ;

[0028] Calculate the normalized comprehensive importance index equal-weighted average score to obtain the final ranking of features. Select the top-ranked core features to form an explicit waveform feature subset. The formula for calculating the equal-weighted average score is as follows:

[0029] ;

[0030] Where M represents the total number of evaluation criteria.

[0031] Based on the descending ranking of equal-weighted average scores, an incremental forward search strategy is adopted to add features sequentially using the average accuracy of the multi-classifiers as a utility benchmark, thereby determining the optimal feature subset dimension.

[0032] Optionally, the implicit deep feature extraction branch in step S3 includes a multi-scale convolution module and a channel attention module cascaded in sequence.

[0033] The multi-scale convolution module has three convolutional branches with different receptive fields set in parallel, and its output is represented as follows:

[0034] ;

[0035] in, This indicates that the kernel size is k. i The sequence of convolution, batch normalization, and nonlinear activation operations, k1, k2, and k3 correspond to the short-time, medium-time, and long-time receptive fields, respectively. This indicates feature concatenation along the channel dimension;

[0036] The channel attention module obtains the channel statistics vector z through global average pooling, and calculates the activation weight vector s through a fully connected dimensionality reduction and dimensionality increase gating mechanism:

[0037] ;

[0038] Where z is the global channel statistics vector, and W1 and W2 are weight matrices. Let Gaussian error be the activation function of the linear unit. The weight is the Sigmoid function, which is used to recalibrate the feature channels to suppress ambient noise.

[0039] Optionally, the implicit deep feature extraction branch, after the multi-scale convolution module and the channel attention module, is further cascaded with a depthwise separable convolutional layer and a residual connection module, and outputs a fixed-length implicit deep feature vector through global average pooling.

[0040] Optionally, the explicit waveform feature transformation branch in step S3 includes a heterogeneous feature mapping layer, which nonlinearly maps the target explicit waveform feature subset to a high-dimensional semantic space through a cascaded combination of batch normalization, fully connected layers, nonlinear activation functions, and regularization layers. The forward propagation formula is:

[0041] ;

[0042] Among them, V man The input is the explicit waveform feature vector, BN is the batch normalization operation, and W is the input waveform feature vector. fc The weight matrix is ​​the fully connected component. It is a non-linear activation function;

[0043] The fused joint feature vector in step S4 is represented as follows:

[0044] ;

[0045] Among them, F deep For implicit depth features, F man The explicit waveform features after mapping, and the fused F fusion This is the joint feature vector.

[0046] Optionally, the adaptive focus loss function used in step S4 introduces an asymmetric penalty mechanism and dynamic class weights, and its calculation method is as follows:

[0047] ;

[0048] in, The predicted probability of the true class. Focusing parameters; indicator functions The value is 1 when the true label of the current sample is background noise, and 0 otherwise; λ is the asymmetric penalty coefficient. The output is the probability of rock falling, used to strictly suppress the probability that the system will falsely report harmless noise as a high-risk event; These are dynamic class weights, adaptively calculated based on sample frequency.

[0049] Optionally, the network model training in step S4 adopts a hybrid decaying learning rate scheduling strategy based on the AdamW optimizer. Linear preheating is used in the initial stage, and a cosine annealing strategy is used after preheating, as follows:

[0050] ;

[0051] in, and These are the maximum and minimum learning rates, respectively. For the current annealing cycle, This refers to the total number of annealing cycles.

[0052] Accordingly, the present invention also provides a system based on the above method, comprising:

[0053] The signal acquisition module is used to deploy a distributed fiber optic acoustic wave sensing system in an underground tunnel, including a sensing optical cable laid along the side wall of the tunnel and a distributed fiber optic acoustic wave sensing host connected to the sensing optical cable at the tunnel exit. The distributed fiber optic acoustic wave sensing system collects acoustic signals and extracts time series signals of preset events.

[0054] The target explicit waveform feature subset acquisition module is used to extract a multi-dimensional waveform feature pool covering the time domain and wavelet packet time-frequency domain features of the time series signal, and to use a multi-criteria feature filtering method to filter and reduce the dimensionality of each feature in the multi-dimensional waveform feature pool to obtain the target explicit waveform feature subset.

[0055] A dual-branch feature extraction network construction module is used to construct a dual-branch feature extraction network comprising an implicit deep feature extraction branch and an explicit waveform feature transformation branch. The implicit deep feature extraction branch receives the time-series signal, extracts temporal features at different time-frequency scales in parallel through a multi-scale convolution module, and adaptively weights the feature channels using a channel attention module to suppress environmental noise, outputting implicit deep features. The explicit waveform feature transformation branch receives the target explicit waveform feature subset, nonlinearly maps the target explicit waveform feature subset to a high-dimensional semantic space through a feature mapping layer, and outputs the mapped explicit waveform features.

[0056] The network model training module is used to cascade and fuse the implicit deep features and the mapped explicit waveform features in the channel dimension into a joint feature vector, and combine the adaptive focus loss function and learning rate scheduling strategy to complete the network model training.

[0057] The early warning module is used to input real-time acoustic signals to be tested into the trained network model and output the category probability of each preset event to realize the classification and early warning of underground tunnel intrusion events.

[0058] Accordingly, the present invention also provides a distributed fiber optic acoustic wave sensing system, including a sensing optical cable laid along the side wall of the tunnel, and a distributed fiber optic acoustic wave sensing host laid at the tunnel exit end and connected to the sensing optical cable.

[0059] The distributed fiber optic acoustic sensor host uses the above method to classify and warn of intrusion events in underground tunnels.

[0060] Accordingly, the present invention also provides a computer terminal, including at least one memory and at least one processor, wherein the memory stores one or more computer instructions, and the one or more computer instructions are executed by the processor to implement the underground tunnel intrusion event identification method provided by the present invention.

[0061] This invention provides a method for identifying intrusion events in underground tunnels, applicable to complex underground engineering environments such as abandoned mine tunnels. This invention utilizes a multi-criteria feature screening framework to comprehensively evaluate, screen, and reduce the dimensionality of the initial waveform feature pool by integrating linear discriminant analysis, nonlinear correlation, and tree model contribution. This effectively eliminates highly collinear redundant features, achieving efficient allocation of computational resources in online monitoring systems. Furthermore, by constructing an implicit deep branch containing multi-scale convolution and channel attention modules, and combining it with an explicit waveform feature branch after heterogeneous mapping for parallel processing, this solves the decision bias problem in classifying highly similar impact signals in conventional deep learning models due to the lack of waveform feature constraints. Finally, network optimization is performed by combining dynamic class weights and asymmetric penalty loss functions with a cosine annealing strategy, making the basic data for locating the objective function more accurate and effective, significantly reducing the false alarm rate of the system. This invention easily obtains the global optimum, exhibits high recognition accuracy, robustness, and good engineering stability. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating the overall method steps of the present invention;

[0063] Figure 2 This is a schematic diagram of the on-site layout of the monitoring system of the present invention;

[0064] Figure 3 This is a flowchart of the Multi-Criterion Feature Filtering Framework (MCFS);

[0065] Figure 4 This is the overall architecture diagram of the dual-branch feature fusion extraction network;

[0066] Figure 5 This is a schematic diagram of the multi-scale convolution module structure for the implicit deep feature extraction branch;

[0067] Figure 6 This is the flowchart of the channel attention module processing;

[0068] Figure 7 This is a schematic diagram of waveform feature heterogeneous mapping;

[0069] Figure 8 This is a performance comparison chart of the method of the present invention and the comparative method under field working conditions. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. Rather, these embodiments are provided to make the invention more thorough and complete, and to fully convey the scope of the invention to those skilled in the art. Furthermore, the technical features involved in the various embodiments of the invention described herein can be combined with each other as long as they do not conflict with each other.

[0071] It should be noted that although functional modules are divided in the system and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the functional module division or the order shown in the flowchart. The terms "first," "second," etc., used in the specification, claims, and drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Unless otherwise defined, all technical and scientific terms used herein have the meanings commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used herein is only for describing embodiments of this application and is not intended to limit this application.

[0072] As can be seen from the background technology, Wu et al. (2021) proposed a coupled architecture that extracts multiple waveform features from the time-frequency domain and connects them to the deep features of a one-dimensional convolutional neural network, and then uses variance analysis and maximum information coefficient for feature selection. However, this method only uses two evaluation criteria, which is difficult to fully cover the linear and nonlinear feature relationships, and it has not been adapted and optimized for the complex acoustic environment of mine roadways. Shi et al. (2023) directly input the original spatiotemporal matrix into the convolutional neural network for end-to-end classification. Although this reduces the amount of manual work, this method lacks explicit waveform feature guidance and has a recognition bottleneck in distinguishing highly similar transient signals. Wu Shihai et al. (2023) proposed a dual-input convolutional neural network that extracts features from the one-dimensional signal and the wavelet time-frequency map respectively and then fuses them for classification, but its fusion strategy does not address the alignment mapping problem of heterogeneous feature spaces. None of the above existing technologies have effectively solved the technical problem that highly similar transient impact signals are easily confused in the environment of abandoned mine roadways.

[0073] To address the aforementioned problems, this invention discloses a method and system for identifying intrusion events in underground tunnels, along with a computer terminal, applicable to complex underground engineering environments such as abandoned mine tunnels. This invention utilizes a multi-criteria feature screening framework to comprehensively evaluate, screen, and reduce the dimensionality of the initial waveform feature pool by integrating linear discriminant analysis, nonlinear correlation, and tree model contribution. This effectively eliminates highly collinear redundant features, achieving efficient allocation of computing resources in the online monitoring system. Furthermore, by constructing an implicit deep branch containing multi-scale convolution and channel attention modules, and coordinating it with an explicit waveform feature branch after heterogeneous mapping for parallel processing, this invention solves the decision bias problem in classifying highly similar impact signals in conventional deep learning models due to the lack of waveform feature constraints. Finally, network optimization is performed by combining dynamic class weights and asymmetric penalty loss functions with a cosine annealing strategy, making the basic data for locating the objective function more accurate and effective, significantly reducing the false alarm rate of the system. This invention easily obtains the global optimum, exhibits high recognition accuracy, robustness, and good engineering stability.

[0074] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0075] Step S1: Deploy a distributed fiber optic acoustic wave sensing system in the underground tunnel, including a sensing optical cable laid along the tunnel sidewall and a distributed fiber optic acoustic wave sensing host connected to the sensing optical cable at the tunnel exit. Acoustic signals are collected and time-series signals of preset events are captured through the distributed fiber optic acoustic wave sensing system.

[0076] Specifically, armored sensing optical cables are laid on the sidewalls of abandoned mine roadways. One end of the sensing optical cable is connected to a distributed optical fiber acoustic wave sensor (DAS) host deployed at the roadway exit. The host emits probe light pulses to the sensing optical cable and receives backscattered signals, thereby realizing continuous acoustic wave monitoring of the entire length of the roadway and collecting acoustic signals.

[0077] In some embodiments, the sampling frequency is set to 100 Hz, and the collected continuous acoustic signal is segmented according to preset events.

[0078] As an example, in this embodiment, the preset events include four types of events: background noise, manual knocking, water seepage, and rockfall. Specifically, each signal segment is truncated for 10 seconds, corresponding to a time series of L=1000 sampling points. The system acquires four types of discrete time series signals, including background noise, artificial knocking, seepage release, and rock falling. These are the acoustic signals of the 1st, 2nd, and Lth sampling points in the time series X. It should also be understood that this embodiment uses the above four types of events as preset events. In other feasible embodiments, adaptive adjustments to the types of preset events based on the application environment and application object also fall within the scope of protection of this invention.

[0079] It should be noted that, in terms of optical cable selection, considering the complex environment of abandoned mine tunnels, armored optical cables with mechanical protection capabilities are given priority. This embodiment conducts systematic comparative tests on various armored optical cables in terms of signal strength, signal-to-noise ratio, time-frequency domain characteristics, and relative attenuation characteristics, selecting the armored optical cable with the best overall performance as the sensing medium to ensure project feasibility and signal quality.

[0080] Step S2: Extract a multi-dimensional waveform feature pool from the time series signal, which includes features in the time domain and wavelet packet time-frequency domain. Then, use the Multi-Criteria Feature Selection (MCFS) method to filter and reduce the dimensionality of each feature in the multi-dimensional waveform feature pool to obtain the target explicit waveform feature subset.

[0081] Specifically, in some embodiments, the following feature extraction and screening process is uniformly performed on the four types of signals collected (background noise, manual knocking, seepage release, and rock falling).

[0082] First, multidimensional waveform features of the original acoustic signal are extracted, including time-domain and time-frequency domain features. For time-domain features, 16 statistical measures are selected, covering energy and amplitude (including peak value, minimum value, peak-to-peak value, mean, rectified average, root mean square value, root square amplitude, and energy, totaling 8 measures), dispersion and statistical distribution (including variance, standard deviation, kurtosis, and skewness, totaling 4 measures), and dimensionless factors (including margin factor, crest factor, impulse factor, and waveform factor, totaling 4 measures). For time-frequency domain features, the original signal is orthogonally divided into four sub-bands: Low-Low Pass (LLP), Low-High Pass (LHP), High-Low Pass (HLP), and High-High Pass (HHP). Six features are extracted for each sub-band, including energy percentage, wavelet entropy, and wavelet information content. The aforementioned time-domain and time-frequency domain features together constitute a 22-dimensional initial waveform feature pool.

[0083] Subsequently, to eliminate the inductive bias of a single evaluation criterion, this embodiment uses the MCFS framework for dimensionality reduction: the F-value is calculated using Analysis of Variance (ANOVA) (Formula (1)) to capture energy class features with strong linear discriminative power, and the linear discriminative power of each feature for the classification label of the preset event is quantified for linear category evaluation. Formula (1) is as follows:

[0084] (1)

[0085] Where K is the number of preset event categories, and N is the total number of samples. Let be the number of samples in class i. Let be the feature mean of the i-th class. This is the population mean; represents the feature value of sample j in class i; MSB (Mean Square Between) is the mean square between groups; MSW (Mean Square Within) is the mean square within groups.

[0086] As an example, in this embodiment, the preset events are divided into four categories: background noise, manual knocking, seepage release, and rock falling, i.e., K=4, and 4500 samples are taken, i.e., N=4500.

[0087] Using mutual information (MI) (Formula (2)) to quantify waveform features X to determine classification labels The provided information gain is used to extract an index from the impact signal that has a complex nonlinear mapping with the event category. Formula (2) is as follows:

[0088] (2)

[0089] Where X is a waveform feature to be evaluated, and Y is an event category label (for example, in this embodiment, the values ​​are 0, 1, 2, and 3, which correspond to four types of events: background noise, manual knocking, seepage release, and rock falling, respectively). It is the marginal entropy of the category. It is conditional entropy. It is a joint probability distribution. and It is a marginal probability distribution, where x is the feature value and y is the class value.

[0090] Global contribution evaluation of tree model: For the above 22-dimensional multi-dimensional waveform feature pool, this invention introduces a tree model algorithm to select features and obtain the contribution as one of the evaluation indicators. It should be understood that in the field of feature selection technology, tree model is used for feature selection of multi-dimensional feature pool and is a mature existing technology. Common algorithms include random forest, XGBoost, gradient boosting tree, etc. This invention will not describe the implementation process in detail. Refer to the existing technology. The purpose of this invention is to introduce this mature algorithm to obtain the global contribution of tree model as an evaluation indicator to improve the multi-dimensional evaluation system. In this embodiment, embedded methods such as random forest (RF) (formula (3)) and extreme gradient boosting model (XGBoost) (formula (4)) are used to calculate the contribution of each waveform feature in the integrated decision and directly quantify the actual role of dimensionless parameters in the complex decision boundary from the perspective of the global contribution of the model.

[0091] Specifically, in this embodiment, the importance of random forest is quantified by calculating the decrease in Gini impurity brought about by waveform features in all decision trees, as shown in formula (3) below:

[0092] (3)

[0093] in, X is the number of trees. j Let b be the j-th waveform feature in the initial 22-dimensional waveform feature pool, and let T be the index of a single decision tree. b Let b be the set of all nodes of the b-th tree. It is a node The splitting characteristics, Is the node reached The sample proportion This is the decrease in the Gini index caused by the split at this node.

[0094] Specifically, in this embodiment, for the XGBoost model, the gain is calculated using objective function optimization, and formula (4) is as follows:

[0095] (4)

[0096] in, and It is the sum of the first derivatives of the left and right child nodes. and It is the sum of the second derivatives. and This is the regularization parameter.

[0097] Finally, the original scores of the above four criteria are normalized using the Min-Max method and all scores are uniformly mapped to the interval [0,1] using formula (5), as follows:

[0098] (5)

[0099] in, For the first One feature in The original importance scores (i.e., evaluation indicator scores) under the criteria correspond to the following: , , and In other words, for each of the 22 features in the initial waveform feature pool, each criterion gives an original score for each feature.

[0100] Then, the normalized comprehensive importance index (NCII), i.e., the equally weighted average score, is calculated according to formula (6) to obtain the feature comprehensive ranking. Formula (6) is as follows:

[0101] (6)

[0102] Where M is the total number of evaluation criteria. In this embodiment, M is 4. After the calculation is completed, each feature gets an NCII value, for a total of 22 NCII values, which are then ranked in descending order.

[0103] Based on the descending ranking of NCII, an incremental forward search strategy is adopted. The feature ranked first is used as the initial subset, and the second-ranked features are added sequentially. After each addition, cross-validation is performed using the average accuracy (Acc) of four classifiers: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), and Random Forest (RF) as the utility benchmark. The average accuracy under each dimension is recorded. When the average accuracy reaches the global maximum value (i.e., BestAcc = Acc), the algorithm is called. k Stop adding when the cutoff point is reached. The dimension corresponding to this cutoff point is the dimension of the best feature subset, and the corresponding feature subset is the target explicit waveform feature subset.

[0104] In this embodiment, as Figure 3As shown, the dimension is 8 dimensions. The optimal feature subset (i.e., the target explicit waveform feature subset) is the final preferred Top-8 explicit waveform feature subset, which includes LHP energy, rectified average value, peak-to-peak value, waveform factor, kurtosis, root amplitude, margin factor and pulse factor.

[0105] Step S3 is executed: constructing a two-branch feature extraction network comprising an implicit deep feature extraction branch and an explicit waveform feature transformation branch; wherein the implicit deep feature extraction branch receives the time series signal, extracts time series features of different time-frequency scales in parallel through a multi-scale convolution module, and performs adaptive weighting on feature channels through a channel attention module to suppress environmental noise, so as to output implicit deep features; the explicit waveform feature transformation branch receives the target explicit waveform feature subset obtained in step S2, nonlinearly maps the target explicit waveform feature subset to a high-dimensional semantic space through a feature mapping layer, and outputs the mapped explicit waveform features.

[0106] Specifically, in some embodiments, an Attention-enhanced Distributed Acoustic Sensing Network (Attentive-DASNet) two-branch network is constructed.

[0107] (1) Implicit deep feature extraction branch:

[0108] Different from the waveform features extracted in step S2, the implicit deep feature extraction branch directly receives the one-dimensional DAS signal, and first automatically learns and extracts time series features of different time-frequency scales in parallel through the multi-scale convolution module. This module is provided with three parallel convolution branches, with convolution kernel sizes of k1, k2 and k3 respectively (k1 < k2 < k3), wherein the short-time branch (k1) has an extremely small receptive field for capturing transient mutation extreme values and high-frequency oscillation details in the signal; the medium-time branch (k2) has a moderate receptive field for detecting envelope changes on the local time scale; the long-time branch (k3) has the largest receptive field for characterizing the macroscopic envelope and low-frequency evolution trend of the signal. The outputs of the three branches are concatenated and spliced on the channel dimension according to formula (7), and formula (7) is as follows:

[0109] (7)

[0110] wherein, is the input one-dimensional signal vector with a length of L, F ki (⋅) represents the operation sequence of convolution, batch normalization and nonlinear activation with the convolution kernel size of k i , and k1, k2 and k3 correspond to the short-time, medium-time and long-time receptive fields respectively, represents performing feature concatenation on the channel dimension, Y msThe temporal feature map output by the multi-scale convolutional module contains multi-scale information extracted by three branches under different receptive fields.

[0111] Subsequently, a squeeze-and-excitation (SE) channel attention module is introduced for adaptive weighting to suppress environmental noise and output implicit deep features.

[0112] First, the C×L dimensional feature map Y output by the multi-scale convolution module is... ms Global average pooling is performed along the time dimension to obtain a C-dimensional global channel statistics vector z, where C is the number of feature channels, and the c-th channel corresponds to the channel statistics vector: u c (i) represents the feature value of the c-th channel at time step i. Subsequently, z is fed into a two-layer fully connected gating mechanism, and the activation weight vector s is calculated according to formula (8). The first fully connected layer compresses the channel dimension from C to C / r, where r is the dimension reduction scaling factor. After activation by a Gaussian Error Linear Unit (GELU), the second fully connected layer restores the dimension to C. Finally, the Sigmoid function maps the channel weights to the [0, 1] interval. Formula (8) is as follows:

[0113] (8)

[0114] Where z is the global channel statistics vector, and W1 and W2 are weight matrices. For GELU activation function, The sigmoid function is used, and this weight is ultimately used to recalibrate the feature channels to suppress environmental noise. This yields the activation weight vector. Then, the weight s of the c-th incentive channel is... c Compared with the c-th channel feature u of the original feature map c Element-wise multiplication completes feature recalibration. For channels characterizing environmental noise floor, s c Approaching 0, adaptive suppression is achieved; for channels characterizing target intrusion events, s c Approaching 1 enables adaptive enhancement.

[0115] Following the attention module, depthwise separable convolutional layers and residual connection modules are further cascaded. Depthwise separable convolution decomposes standard convolution into two steps: channel-wise convolution and pointwise convolution, reducing the number of parameters and computational complexity while maintaining the receptive field. Residual connections directly transfer shallow features to deeper layers through identity mapping, mitigating the degradation problem in deep networks. Max pooling layers are then applied after each convolutional module to progressively reduce the temporal dimension, and dropout layers with varying ratios are cascaded at different depths to prevent overfitting. Finally, adaptive global average pooling compresses the feature map into a fixed-length high-dimensional implicit feature vector F. deep This refers to implicit deep features.

[0116] In a specific example, the complete data flow of the implicit deep feature extraction branch is: input raw signal X → multi-scale convolution module output Y. ms (32 channels) → SE attention module recalibration → Depthwise separable convolution (32 channels → 64 channels) → Residual connection module (64 channels → 128 channels) → Max pooling layer to reduce temporal dimension → Adaptive global average pooling to compress the temporal dimension to 1 → Flattened into a 128-dimensional vector → Channel attention weighted layer → Output 128-dimensional implicit deep feature vector F deep .

[0117] (2) Explicit waveform feature transformation branch:

[0118] The target explicit waveform feature subset obtained in the explicit waveform feature transformation branch receiving step S2 is used as the input vector V. man The heterogeneous feature mapping layer of formula (9) is used for processing, as follows:

[0119] (9)

[0120] Among them, V man The input is the explicit waveform feature vector, BN is the batch normalization operation, and W is the input waveform feature vector. fc The weight matrix is ​​the fully connected component. is a non-linear activation function, and b is the bias vector of the fully connected layer.

[0121] This mapping layer consists of a cascaded batch normalization layer, a fully connected layer, a non-linear activation function, and a Dropout regularization layer, achieving two core functions: first, semantic space alignment, achieved through a learnable weight matrix W. fc The approach involves two main methods: first, mapping low-dimensional waveform features to a high-dimensional representation space similar to deep semantics; and second, nonlinear combinatorial mining, which uses nonlinear activation functions to deeply mine higher-order coupling relationships between explicit waveform features. A feature mapping layer nonlinearly maps a subset of target explicit waveform features to a high-dimensional semantic space, outputting the mapped explicit waveform feature F. man .

[0122] Step S4: Concatenate and fuse the implicit deep features and the mapped explicit waveform features in the channel dimension to form a joint feature vector, and combine the adaptive focus loss function and learning rate scheduling strategy to complete the network model training.

[0123] Specifically, the F output of the implicit depth branch deep F with explicit waveform branch output man According to formula (10), the features are cascaded and fused into a joint feature vector F along the channel dimension. fusion Formula (10) is as follows:

[0124] (10)

[0125] Among them, F deep For implicit deep features, F man The explicit waveform features after mapping, and the fused F fusion This is the joint feature vector.

[0126] This fusion strategy preserves the independent boundaries of explicit waveform features to the greatest extent possible, enabling the model to both uncover deep implicit distribution features of the data and make joint decisions by combining explicit statistical indicators. The fused features are then fed into a softmax classifier after a dropout layer to output the predicted probabilities of each event category.

[0127] Next, the network model is trained by combining the adaptive focus loss function with the learning rate scheduling strategy. The loss function adopted is the adaptive focus loss function described in formula (11), which is as follows:

[0128] (11)

[0129] Where, p t γ is the predicted probability of the Softmax classification layer of the model for the true class of the current sample (i.e., the probability value of class t output by Softmax when the true label of the sample is class t), γ is the focusing parameter, and γ is the indicator function. The value is 1 when the true label of the current sample is background noise, and 0 otherwise. λ is the asymmetric penalty coefficient. The output probability of a rock falling is used to strictly suppress the probability that the system will falsely report harmless noise as a high-risk event. These are dynamic class weights, adaptively calculated based on sample frequency.

[0130] This function introduces a focusing parameter γ to dynamically decay the loss weights of easily distinguishable samples, guiding the model to shift its optimization focus to the difficult-to-distinguish boundary samples. Simultaneously, the class balance factor α... tAdaptive calculation based on the sample frequency of each category is used to mitigate the objective differences in the frequency of various intrusion events in the dataset. Furthermore, an asymmetric penalty mechanism is designed: when the true label is background noise, a penalty is applied through an indicator function. An additional penalty term proportional to the predicted rockfall probability is applied to strictly suppress the probability of the system misreporting harmless noise as a high-risk rockfall event at the objective function level.

[0131] The optimizer used is AdamW, and the weight decay factor is set to 1×10. -4 The learning rate scheduling adopts a hybrid decay strategy combining linear warm-up and cosine annealing: in the warm-up phase of the initial training, the learning rate linearly increases from 0 to the initial peak; after the warm-up, cosine annealing is performed according to formula (12), and the learning rate smoothly falls back along the cosine curve, ensuring that the model explores and converges to a better solution with a small step size in the later stage of training. Formula (12) is as follows:

[0132] (12)

[0133] in, and These are the maximum and minimum learning rates, respectively. For the current annealing cycle, This refers to the total number of annealing cycles.

[0134] Specifically, the training process of the network model is as follows: First, the training set samples are fed into the network for forward propagation in batches of 64 to obtain the class probability distribution of the Softmax output; then, the adaptive focus loss L of the batch is calculated according to formula (11). total Next, the gradient of the loss with respect to each network parameter is calculated through backpropagation; finally, the AdamW optimizer updates the network weights based on the current learning rate (determined by formula (12)) and the gradient direction. Specifically, the above process of forward propagation → loss calculation → backpropagation → parameter update is executed cyclically until the maximum number of training rounds is reached or the early stopping condition is triggered.

[0135] Data augmentation strategies are introduced during the training phase, including random amplitude scaling, frequency domain Gaussian noise addition, and time sequence flipping.

[0136] Step S5: Input the real-time acoustic signal to be tested into the trained network model, and output the category probability of each preset event to realize the classification and early warning of underground tunnel intrusion events.

[0137] Specifically, the trained model is deployed at the edge of the site, receives data streams collected by the DAS system in real time for online prediction, and executes safety, early warning or alarm commands based on the highest probability category output by the Softmax classification layer.

[0138] The following is in conjunction with the appendix Figure 1 To be continued Figure 8 This paper provides a detailed explanation of the specific implementation effects of the present invention in the monitoring of intrusion events in legacy underground tunnels.

[0139] (1) Construction of experimental platform and dataset.

[0140] Figure 2 This is a schematic diagram of the on-site layout of the monitoring system of the present invention. Figure 2 As shown, the monitoring system mainly consists of a distributed optical fiber acoustic sensor (DAS) host and sensor optical cables. The core monitoring medium is preferably armored sensor optical cables, which are continuously laid along the sidewalls of abandoned mine roadways and securely fixed to the roadway walls using nail-free adhesive and saddle clips. One end of the optical cable is connected to the DAS host, utilizing the Rayleigh backscattering principle of optical fiber to achieve continuous monitoring along the entire line. During long-distance deployment, cable reeling points can be set according to actual project needs, i.e., a certain length of cable is reserved at specific locations to enhance local spatial sampling density and facilitate later maintenance and adjustments.

[0141] In the specific implementation and verification phase, this embodiment first constructed a test platform simulating the environment of abandoned mine tunnels in the air-raid shelter of Area A of Chongqing University. Through systematic comparative testing of five types of optical cables (one type of unarmored optical cable and four types of armored optical cables with different sheath structures) in terms of signal strength, signal-to-noise ratio, relative attenuation, and mid-to-high frequency ratio, the results showed that armored optical cable 4 had the best overall performance: its relative attenuation was only 5.05 dB, its signal-to-noise ratio reached 10.55 dB, and its mid-to-high frequency ratio reached 84.73%, maintaining good acoustic transmission performance and high-frequency enhancement capabilities while providing mechanical protection.

[0142] Based on this, acoustic signals of four typical events—background noise, manual knocking, water seepage, and rockfall—were designed and collected. After manual truncation and label verification, each segment was L=1000 sampling points (corresponding to 10 seconds), constructing an acoustic dataset of 4500 samples (i.e., 4500 10-second signal segments) of mine legacy roadway intrusion events. A hierarchical random partitioning method was used to divide the dataset into a training set (3600 samples) and a validation set (900 samples) at an 8:2 ratio, ensuring a balanced distribution of samples across the four event categories.

[0143] (2) Multi-criteria feature screening results.

[0144] Figure 3This paper demonstrates the complete process of the MCFS multi-criteria feature selection framework. In this embodiment, the initial waveform feature pool contains 16 time-domain features and 6 time-frequency domain features, totaling 22 dimensions. Through comprehensive evaluation using four complementary criteria—ANOVA, MI, RF, and XGBoost—the original scores of each criterion are normalized, and the NCII is calculated and ranked in descending order. Incremental forward search validation shows that when the feature dimension is reduced from 22 to 8 dimensions (a reduction of approximately 63.6%), the average accuracy of the four classifiers reaches the highest global value. The final selected Top-8 core feature subset includes: LHP energy, rectified average value, peak-to-peak value, waveform factor, kurtosis, root square magnitude, margin factor, and impulse factor.

[0145] Visual comparison using the t-SNE algorithm revealed that the intra-class distribution of various events in the original 22-dimensional feature space was relatively dispersed, with significant overlap between rockfall and manual knocking sample clusters. After dimensionality reduction to 8 dimensions using MCFS, the intra-class samples of each category clustered more tightly, the inter-class distances increased significantly, and the clustering effect was significantly improved. Machine learning cross-validation further confirmed that the optimized Top-8 feature subset improved the accuracy of KNN from 93.56% to 95.89%, the accuracy of decision trees from 94.33% to 95.78%, and the accuracy of random forest remained unchanged at 96.44%, indicating that the core feature information was intact and the reduction in dimensionality did not result in the loss of key features.

[0146] (3) Performance verification of the Attentive-DASNet model.

[0147] Figure 4 The overall architecture of the Attentive-DASNet dual-branch feature fusion network is shown. Figure 5 The structure of the multi-scale convolution module is shown, where three parallel convolutions with k1 = 3, k2 = 15, and k3 = 25 capture local transient details, mid-frequency rhythms, and global low-frequency trends, respectively. Figure 6 The processing flow of the channel attention module is demonstrated, which achieves feature-level "soft noise reduction" through global compression and adaptive excitation. Figure 7 A schematic diagram of waveform feature heterogeneous mapping is shown, which maps low-dimensional waveform features to a high-dimensional space to achieve semantic alignment.

[0148] In this embodiment, the model input signal length L = 1000, and the optimizer selected is AdamW (initial learning rate 1×10⁻⁶). −4 Weight decay 1×10 −4 The batch size is 64, the maximum number of training rounds is 300 (including 20 rounds of linear warm-up), the focus parameter γ = 2 in the focus loss function, and the asymmetric penalty coefficient λ = 0.1.

[0149] The classification performance of each model on the indoor validation set is compared in Table 1 below:

[0150] Table 1. Comparison of classification performance of each model

[0151]

[0152] As shown in Table 1, the method of the present invention achieves an accuracy of 98.11% on the indoor validation set, which is 1.33% higher than that of pure CNN, and the average loss is 45.5% lower than that of pure CNN. This shows that the explicit waveform features provide effective physical constraints for the network, significantly improving the prediction confidence and decision boundary clarity.

[0153] Table 2 Performance of Fusion Models by Category

[0154]

[0155] As shown in Table 2, in distinguishing between rockfalls and human-induced knocking events, which are the most challenging aspects, the precision of the rockfall category improved from 91.05% for pure CNN to 95.88% (an increase of 4.83%), and the recall improved to 98.73%. Analysis of the confusion matrix revealed that pure CNN misclassified human-induced knocking as rockfalls in 15 cases, while the method of this invention significantly reduced this misclassification to 4 cases by introducing explicit waveform feature constraints such as kurtosis and impulse factor.

[0156] By visually comparing the high-dimensional feature vectors before the fully connected classification layer using the t-SNE algorithm, it was found that the feature distribution of the pure CNN model lacked clear clustering boundaries between different categories of events, and there was a serious overlap between rockfall and manual knocking sample points. However, in the feature distribution of the method of this invention, the previously seriously overlapping rockfall and manual knocking sample points were effectively separated, forming a clearer classification boundary. This demonstrates that the explicit waveform features compensate for the shortcomings of a single convolution operation in feature discrimination.

[0157] (4) Ablation experiment.

[0158] To verify the independent contributions of each core module, the following ablation experiments were conducted: After removing the SE channel attention module, the model's validation set accuracy decreased from 98.11% to 95.89%, a performance loss of 2.22%. The recall rate for manual knocking decreased to 90.62%, with approximately 9.3% of knocking samples misclassified as falling rocks, indicating that the SE module plays a crucial role in suppressing background noise and distinguishing highly similar transient signals. However, after removing the waveform feature branch (i.e., pure CNN), the average validation set loss increased from 0.0073 to 0.0134, an increase of approximately 84%, demonstrating that explicit waveform features provide additional orthogonal gradient directions for network optimization during backpropagation, enabling the model to form classification boundaries with larger intervals.

[0159] (5) Comparison of multi-scale convolution kernel combinations.

[0160] This example compares the performance differences between three single-size convolutional kernels and multi-scale parallel architectures, as shown in the table below:

[0161] As shown in Table 3, the multi-scale parallel architecture can simultaneously extract transient high-frequency mutations and long-term low-frequency envelope features in a single forward propagation, achieving a validation set accuracy of 98.11%, which is better than all single-scale models.

[0162] Table 3 Performance Comparison of Different Convolutional Kernel Combinations

[0163]

[0164] (6) On-site engineering verification.

[0165] To verify the model's engineering applicability under real-world complex conditions, a field deployment and test of the DAS system was conducted in a non-metallic mine roadway. The fiber optic cable deployment type was consistent with the indoor test, using armored fiber optic cables with optimal overall performance. The field test collected continuous monitoring data for four types of events. After manual annotation and outlier removal, 1200 high-quality field samples (300 per type) were obtained, which were divided into a training set (840 samples) and a test set (360 samples) at a 7:3 ratio. Both the pure CNN and the fusion model were loaded with weights pre-trained on the indoor dataset, using a 1×10⁻⁶ weight distribution. −4 The learning rate is fine-tuned with all parameters.

[0166] On-site test results as follows Figure 8 As shown in Table 3:

[0167] As shown in Table 4, the method disclosed in this invention achieved a maximum accuracy of 97.78% under complex on-site conditions, which is 3.34% and 2.50% higher than the random forest and pure CNN models, respectively, and only 0.33% lower than the test accuracy on the indoor dataset. In the on-site environment, all three models achieved 100% accuracy in identifying background noise. The main classification difficulties were concentrated on two types of transient impact signals: manual knocking and rockfall. The pure CNN model misclassified manual knocking as rockfall in 12 cases, while the method of this invention reduced this misclassification to 6 cases. Furthermore, it did not exhibit cross-contamination with water seepage or other impact events, effectively reducing the false alarm probability of high-risk events in real engineering scenarios.

[0168] Table 4. Field Test Results

[0169]

[0170] The above results demonstrate that the method of the present invention maintains robust recognition performance in real and complex environments, and preliminarily verifies the engineering applicability of the dual-branch fusion scheme in monitoring intrusion events in abandoned mine roadways.

[0171] Accordingly, embodiments of the present invention also provide a method-based system, including a signal acquisition module connected sequentially or interconnected, a target explicit waveform feature subset acquisition module, a dual-branch feature extraction network construction module, a network model training module, and an early warning module.

[0172] The signal acquisition module is used to deploy a distributed fiber optic acoustic wave sensing system in an underground tunnel, including a sensing optical cable laid along the side wall of the tunnel and a distributed fiber optic acoustic wave sensing host connected to the sensing optical cable at the tunnel exit. The distributed fiber optic acoustic wave sensing system collects acoustic signals and extracts time series signals of preset events.

[0173] The target explicit waveform feature subset acquisition module is used to extract a multi-dimensional waveform feature pool covering the time domain and wavelet packet time-frequency domain features of the time series signal, and to use a multi-criteria feature filtering method to filter and reduce the dimensionality of each feature in the multi-dimensional waveform feature pool to obtain the target explicit waveform feature subset.

[0174] A dual-branch feature extraction network construction module is used to construct a dual-branch feature extraction network comprising an implicit deep feature extraction branch and an explicit waveform feature transformation branch. The implicit deep feature extraction branch receives the time-series signal, extracts temporal features at different time-frequency scales in parallel through a multi-scale convolution module, and adaptively weights the feature channels through a channel attention module to suppress environmental noise, outputting implicit deep features. The explicit waveform feature transformation branch receives the target explicit waveform feature subset obtained in step S2, nonlinearly maps the target explicit waveform feature subset to a high-dimensional semantic space through a feature mapping layer, and outputs the mapped explicit waveform features.

[0175] The network model training module is used to cascade and fuse the implicit deep features and the mapped explicit waveform features in the channel dimension into a joint feature vector, and combine the adaptive focus loss function and learning rate scheduling strategy to complete the network model training.

[0176] The early warning module is used to input real-time acoustic signals to be tested into the trained network model and output the category probability of each preset event to realize the classification and early warning of underground tunnel intrusion events.

[0177] Accordingly, the present invention provides a distributed fiber optic acoustic wave sensing system, comprising a sensing optical cable laid along the sidewall of a tunnel, and a distributed fiber optic acoustic wave sensing host connected to the sensing optical cable and laid at the tunnel exit. The host includes a processor that performs time-series signal processing, specifically referring to steps S1-S5 to achieve classification and early warning of underground tunnel intrusion events.

[0178] Accordingly, embodiments of the present invention also provide a computer terminal, including at least one memory and at least one processor. The memory stores one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the above-described method for identifying underground tunnel intrusion events. The specific implementation process can be referred to the method description in the foregoing embodiments, and will not be repeated here.

[0179] Accordingly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which is invoked by a processor to implement the above-described method for identifying underground tunnel intrusion events. The specific implementation process can be referred to the method description in the foregoing embodiments, and will not be repeated here.

[0180] In the description of this application, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0181] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0182] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0183] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0184] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0185] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0186] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

[0187] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

Claims

1. A method for identifying intrusion events in underground tunnels, characterized in that, The method includes the following steps: Step S1: Deploy a distributed fiber optic acoustic wave sensing system in the underground tunnel, including a sensing optical cable laid along the tunnel sidewall and a distributed fiber optic acoustic wave sensing host connected to the sensing optical cable at the tunnel exit. Acoustic signals are collected and time-series signals of preset events are captured through the distributed fiber optic acoustic wave sensing system. Step S2: Extract a multi-dimensional waveform feature pool covering the time domain and wavelet packet time-frequency domain features from the time series signal, and use a multi-criteria feature filtering method to filter and reduce the dimensionality of each feature in the multi-dimensional waveform feature pool to obtain a target explicit waveform feature subset; Step S3: Construct a dual-branch feature extraction network comprising an implicit deep feature extraction branch and an explicit waveform feature transformation branch; wherein, the implicit deep feature extraction branch receives the time series signal, extracts temporal features at different time-frequency scales in parallel through a multi-scale convolution module, and adaptively weights the feature channels through a channel attention module to suppress environmental noise, outputting implicit deep features; the explicit waveform feature transformation branch receives the target explicit waveform feature subset obtained in step S2, nonlinearly maps the target explicit waveform feature subset to a high-dimensional semantic space through a feature mapping layer, and outputs the mapped explicit waveform features; Step S4: Concatenate and fuse the implicit deep features and the mapped explicit waveform features in the channel dimension to form a joint feature vector, and combine the adaptive focus loss function and learning rate scheduling strategy to complete the network model training; Step S5: Input the real-time acoustic signal to be tested into the trained network model, and output the category probability of each preset event to realize the classification and early warning of underground tunnel intrusion events.

2. The method according to claim 1, characterized in that, Step S2 involves using a multi-criteria feature filtering method to filter and reduce the dimensionality of each feature in the multidimensional waveform feature pool, including: First, three complementary evaluation criteria are introduced to evaluate each type of waveform feature, resulting in evaluation index scores. These criteria are: linear discriminant evaluation, nonlinear correlation evaluation, and tree model global contribution evaluation. The linear discriminant evaluation uses analysis of variance to calculate the F-value, serving as the raw importance score for this criterion, quantifying the linear discriminative ability of each feature to classify the preset event's label. The nonlinear correlation evaluation uses mutual information (MI) to quantify the waveform features. To determine event category labels The provided information gain serves as a criterion for the original importance score; the global contribution evaluation of the tree model is calculated by using a random forest and extreme gradient boosting model to determine the contribution of each waveform feature in the ensemble decision. Secondly, minimum-maximum normalization is used to uniformly map the scores of all evaluation indicators to the [0,1] interval; Next, the normalized comprehensive importance index is calculated using an equal-weighted average score to obtain the final ranking of the features. The top-ranked core features are selected to form an explicit waveform feature subset. The formula for calculating the equal-weighted average score is as follows: ； Where M represents the total number of evaluation criteria. The evaluation index score of the i-th waveform feature under evaluation criterion m; Finally, based on the descending ranking of the equally weighted average score, an incremental forward search strategy is adopted to add features sequentially with the average accuracy of the multi-classifier as the utility benchmark, determine the optimal feature subset dimension, and the corresponding feature subset is the target explicit waveform feature subset.

3. The method according to claim 1, characterized in that, The implicit deep feature extraction branch in step S3 includes a multi-scale convolution module and a channel attention module cascaded in sequence. The multi-scale convolution module has three convolutional branches with different receptive fields set in parallel, and its output is represented as follows: ； in, This indicates that the kernel size is k. i The sequence of convolution, batch normalization, and nonlinear activation operations, k1, k2, and k3 correspond to the short-time, medium-time, and long-time receptive fields, respectively. This indicates feature concatenation along the channel dimension, Y ms The temporal feature map output by the multi-scale convolutional module; The channel attention module targets the temporal feature map Y. ms The channel statistics vector z is obtained through global average pooling, and the activation weight vector is calculated through a fully connected dimensionality reduction and dimensionality increase gating mechanism. Obtain the incentive weight vector Then, the time-series feature map Y is analyzed. ms Perform feature recalibration, i.e., the weight s of the c-th excitation channel. c With time series feature map Y ms The c-th channel feature u c Element-by-element multiplication: ； Where z is the global channel statistics vector, and W1 and W2 are weight matrices. Let Gaussian error be the activation function of the linear unit. The sigmoid function is used to recalibrate the feature channels to suppress environmental noise; C is the number of feature channels. This represents the fully connected dimensionality reduction and dimensionality increase gating mechanism, and W represents the general symbol for weights.

4. The method according to claim 3, characterized in that, The implicit deep feature extraction branch, after the multi-scale convolution module and the channel attention module, is further cascaded with a depthwise separable convolutional layer and a residual connection module, and finally outputs a fixed-length implicit deep feature vector through global average pooling.

5. The method according to claim 1, characterized in that, The explicit waveform feature transformation branch in step S3 includes a heterogeneous feature mapping layer, which nonlinearly maps the target explicit waveform feature subset to a high-dimensional semantic space through a cascaded combination of batch normalization, fully connected layers, nonlinear activation functions, and regularization layers. The forward propagation formula is as follows: ； Among them, V man The input is the explicit waveform feature vector, BN is the batch normalization operation, and W is the input waveform feature vector. fc The weight matrix is ​​the fully connected component. It is a non-linear activation function; The fused joint feature vector in step S4 is represented as follows: ； Among them, F deep For implicit depth features, F man The explicit waveform features after mapping, and the fused F fusion This is the joint feature vector.

6. The method according to claim 1, characterized in that, The adaptive focus loss function used in step S4 introduces an asymmetric penalty mechanism and dynamic class weights, and its calculation method is as follows: ； in, The predicted probability of the true class. Focusing parameters; indicator functions The value is 1 when the true label of the current sample is background noise, and 0 otherwise; λ is the asymmetric penalty coefficient. The output is the probability of rock falling, used to strictly suppress the probability that the system will falsely report harmless noise as a high-risk event; These are dynamic class weights, adaptively calculated based on sample frequency.

7. The method according to claim 1, characterized in that, The network model training in step S4 employs a hybrid decaying learning rate scheduling strategy based on the AdamW optimizer. Linear preheating is used in the initial stage, followed by a cosine annealing strategy after preheating, as follows: ； in, and These are the maximum and minimum learning rates, respectively. For the current annealing cycle, This refers to the total number of annealing cycles.

8. A system based on the method of any one of claims 1-7, characterized in that, include: The signal acquisition module is used to deploy a distributed fiber optic acoustic wave sensing system in an underground tunnel, including a sensing optical cable laid along the side wall of the tunnel and a distributed fiber optic acoustic wave sensing host connected to the sensing optical cable at the tunnel exit. The distributed fiber optic acoustic wave sensing system collects acoustic signals and extracts time series signals of preset events. The target explicit waveform feature subset acquisition module is used to extract a multi-dimensional waveform feature pool covering the time domain and wavelet packet time-frequency domain features of the time series signal, and to use a multi-criteria feature filtering method to filter and reduce the dimensionality of each feature in the multi-dimensional waveform feature pool to obtain the target explicit waveform feature subset. A dual-branch feature extraction network construction module is used to construct a dual-branch feature extraction network comprising an implicit deep feature extraction branch and an explicit waveform feature transformation branch. The implicit deep feature extraction branch receives the time-series signal, extracts temporal features at different time-frequency scales in parallel through a multi-scale convolution module, and adaptively weights the feature channels using a channel attention module to suppress environmental noise, outputting implicit deep features. The explicit waveform feature transformation branch receives the target explicit waveform feature subset, nonlinearly maps the target explicit waveform feature subset to a high-dimensional semantic space through a feature mapping layer, and outputs the mapped explicit waveform features. The network model training module is used to cascade and fuse the implicit deep features and the mapped explicit waveform features in the channel dimension into a joint feature vector, and combine the adaptive focus loss function and learning rate scheduling strategy to complete the network model training. The early warning module is used to input real-time acoustic signals to be tested into the trained network model and output the category probability of each preset event to realize the classification and early warning of underground tunnel intrusion events.

9. A distributed fiber optic acoustic wave sensing system, characterized in that: It includes a sensing optical cable laid along the side wall of the tunnel, and a distributed fiber optic acoustic wave sensing host laid at the tunnel exit end and connected to the sensing optical cable. The distributed fiber optic acoustic sensor host uses the method described in any one of claims 1-7 to achieve classification and early warning of underground tunnel intrusion events.

10. A computer terminal, characterized in that, It includes at least one memory and at least one processor, the memory storing one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the underground tunnel intrusion event identification method as described in any one of claims 1-7.

Images